SkillScaffold
A training scaffold method for agent reinforcement learning that estimates skill utility at low cost through outcome-aware likelihood comparison, adaptively selects whether to inject or discard skills based on utility, and jointly trains a skill generator and an executor to promote skill internalization.
Timeline (last 90 days)
SkillScaffold reuses training trajectories whose results have already been verified and uses outcome-aware likelihood comparison to estimate skill utility, without relying at all on candidate-specific environment interaction.
SkillScaffold contains two roles, Generator and Executor, which are respectively responsible for producing skills and executing tasks, and the two are jointly trained.
SkillScaffold's ablation experiments verified that the utility-based selection and abandonment mechanism is one of the key sources of performance improvement.
SkillScaffold was evaluated on four benchmarks—ALFWorld, WebShop, ScienceWorld, and τ²-Bench—and surpassed strong agentic RL baselines and skill-driven RL baselines on both short-horizon and long-horizon tasks.
SkillScaffold's code has been open-sourced on GitHub at https://github.com/Vaidurya00/SkillScaffold.
All Facts (5)
SkillScaffold reuses training trajectories whose results have already been verified and uses outcome-aware likelihood comparison to estimate skill utility, without relying at all on candidate-specific environment interaction.
50%UnverifiedSkillScaffold contains two roles, Generator and Executor, which are respectively responsible for producing skills and executing tasks, and the two are jointly trained.
50%UnverifiedSkillScaffold's ablation experiments verified that the utility-based selection and abandonment mechanism is one of the key sources of performance improvement.
50%UnverifiedSkillScaffold was evaluated on four benchmarks—ALFWorld, WebShop, ScienceWorld, and τ²-Bench—and surpassed strong agentic RL baselines and skill-driven RL baselines on both short-horizon and long-horizon tasks.
50%UnverifiedSkillScaffold's code has been open-sourced on GitHub at https://github.com/Vaidurya00/SkillScaffold.
50%